Jev Language Understanding Embeddings (Qwen3.6-35B-A3B-FP8)

Each split holds state_embeddings.pt, action_embeddings.pt, neg_embeddings.pt, neg_counts.pt, state_options_embeddings.pt, has_options.pt, single_tokens.npz and metadata.json. Tensors are written with torch.save and can be read with torch.load(..., mmap=True); row i of every tensor matches samples[i] in metadata.json.

Qwen/Qwen3.6-35B-A3B-FP8, run through a vLLM pooling runner with last-token pooling and L2 normalization, producing 2048-d vectors. The metadata records max_model_len 2048 and state_head_tokens 1024.

The train splits of iSarcasmEval, VAST, ACOS and New Yorker caption matching, rendered the way the index renders its scored items. The train split has 58,962 pairs across 13,111 tasks, with 18,013 one-token actions and 29,996 one-token negatives.

The split is by task, not by row: a task goes to validation when int(task_id[:8], 16) % 10 == 0. That yields 6,455 validation pairs across 1,466 tasks, with 2,000 one-token actions and 3,331 one-token negatives.

vLLM's pooling of a one-token prompt returns noise for this model, so every one-token action, negative or state takes its vector from the jackyk02/qwen3.6-35b-a3b-fp8-single-token-embeddings table. single_tokens.npz lists those rows and their token ids, and EmbeddingDir applies it on load.

The sources' licenses, listed in the dataset documentation. The vectors are a derivative of the source text and carry its terms, so downstream use inherits the terms of the underlying benchmarks.

The other criteria of each decision are used as hard negatives, four per row (k=4), stored zero-padded in neg_embeddings.pt. neg_counts.pt gives the number of valid negatives for each row, so padded slots can be masked out.